---
library_name: pytorch
license: other
tags:
- glycans
- proteins
- protein-glycan
- affinose
- bertose
- esm-c
- pytorch
---

# AFFINose Interaction Model

This repository contains the AFFINose checkpoint for protein-glycan interaction inference. AFFINose combines BERTose glycan token representations with per-residue ESM-C protein embeddings and returns a scalar interaction score.

## Quick Start

The recommended user path is the companion notebook. For direct Python use, download the checkpoint and vocabulary with `huggingface_hub`:

```python
from huggingface_hub import hf_hub_download

checkpoint = hf_hub_download(
    repo_id="supanthadey1/affinose-interaction-model",
    filename="checkpoints/affinose_interaction_model.pt",
)
vocab = hf_hub_download(
    repo_id="supanthadey1/affinose-interaction-model",
    filename="vocab/bpe_vocabulary.json",
)
```

No Hugging Face token is required for this AFFINose checkpoint now that the repository is public. ESM-C is separate and may require the user's own Hugging Face login depending on EvolutionaryScale access requirements.

## Files

- `checkpoints/affinose_interaction_model.pt` - AFFINose interaction checkpoint.
- `vocab/bpe_vocabulary.json` - WURCS BPE vocabulary for glycan tokenization.
- `src/affinose_model.py` - AFFINose architecture.
- `src/affinose_inference.py` - standalone inference helper.
- `src/affinose_dataset.py` - tokenizer and data utility helpers.
- `src/bertose_model.py` - BERTose model definition used for glycan encoding.
- `src/bertose_layers.py` - Transformer layers used by BERTose.
- `src/wurcs_bpe_tokenizer.py` - WURCS BPE tokenizer.

## Input

Provide one protein-glycan pair or a CSV batch. Glycans should be WURCS strings. Proteins can be provided as IDs linked to precomputed embeddings, or through the companion notebook as raw sequences that are embedded with ESM-C 300M.

Batch CSVs use `sample_id,protein_id,protein_sequence,glycan_wurcs`. Free-text glycan names, common names, SNFG drawings, and IUPAC-condensed strings are not parsed directly by AFFINose. Convert those inputs to WURCS first, then score the protein-glycan pair.

## Protein Embedding Requirement

AFFINose expects per-residue ESM-C 300M embeddings with shape `[L, 960]`. Do not mean-pool the protein before passing it into AFFINose.

ESM-C is a separate EvolutionaryScale protein model. The ESM-C weights are not included in this repository. Users should install the `esm` package and let it download ESM-C 300M into their own runtime cache.

```python
from esm.models.esmc import ESMC
from esm.sdk.api import ESMProtein, LogitsConfig

esmc = ESMC.from_pretrained("esmc_300m").to("cuda")  # or "cpu"
protein = ESMProtein(sequence="MKTAYIAKQRQISFVKSHFSRQLEERLGLIEVQ")
protein_tensor = esmc.encode(protein)
output = esmc.logits(
    protein_tensor,
    LogitsConfig(sequence=True, return_embeddings=True),
)
protein_embeddings = output.embeddings  # per-residue ESM-C 300M embeddings
```

If Hugging Face requests authentication for ESM-C, users should authenticate with their own Hugging Face account/token and accept any required EvolutionaryScale terms. BERTose/AFFINose tokens are not required once these repositories are public.

## Output

A scalar protein-glycan interaction score from the trained AFFINose head.

## Scope

This repository does not perform IUPAC-condensed/name-to-WURCS conversion. For now, provide WURCS directly.

License metadata is currently `other`; update it when the final release license and citation text are chosen.

## References

- EvolutionaryScale ESM package: https://github.com/evolutionaryscale/esm
- ESM-C 300M Hugging Face model: https://huggingface.co/EvolutionaryScale/esmc-300m-2024-12